The expansion of Phyllostachys edulis (Moso Bamboo) into Cunninghamia lanceolata (Chinese fir) forest has led to ecological and economic problems such as “forest retreat and bamboo advance” and forest rights disputes. The use of remote sensing to effectively invert the succession process of Moso bamboo is of great significance for scientific control of forest resources. To reveal the effectiveness of canopy hyperspectral inversion of leaf area index (LAI) at different expansion stages of Moso bamboo to Chinese fir forest, four types of sample squares in mixed forest, which were divided based on a percentage of Moso bamboo were set to simulate expansion stages Ⅰ, Ⅱ, Ⅲ and Ⅳ along the expansion direction. Meanwhile, to explore the applicability of LAI hyperspectral inversion models for different expansion stages of Moso bamboo to Chinese fir forest, five types of single-factor regression models were established based on the characteristic wavebands which were chosen by the correlation between the original spectra, 10 spectral transformations such as open square, logarithmic, inverse, first-order differential and second-order differential and LAI at different expansion stages, and seven LAI significantly related vegetation indices such as normalized difference vegetation index (NDVI), yellowness index (YI) and so on. Multi-factor regression models were established based on the vegetation indices using four machine learning methods: neural network, decision forest regression, Bayesian linear regression, and linear regression. The results showed that the differential transform, including first-order differentiation (R′), second-order differentiation (R″), logarithmic first-order differentiation ［(lgR)′］ and inverse first-order differentiation ［(1/R)′］ of the original spectrum (R), could enrich the spectral information to characterize the expansion process better. Among the vegetation indices, YI had the highest correlation coefficient with LAI, showing high sensitivity to the expansion process, and NDVI had the best inversion effect. However, the overall inversion based on traditional vegetation index modeling was in effective. Moreover, the quadratic polynomial and power exponential regression models based on the differential transform spectra performed better in each expansion stage, while the inversion effect based on the vegetation indices was poor, and the neural network algorithm outperformed other machine learning algorithms. Traditional regression algorithms based on spectral transformations performed better than machine learning modeling methods. The model (y=5.291 4e183.76x) based on log-inverse first-order differential transform spectra fitted best in expansion stage Ⅲ with R2 of 0.735 and 0.742, RMSE of 0.733 and 0.468, and nRMSE of 14.0% and 9.9% for the modeling and validation sets, respectively. We suggest that the Moso bamboo expansion control should be selected in mixed forests of half of each species. Innovative analysis of LAI hyperspectral inversion modeling of Moso bamboo at different expansion stages will provide a basis for scientific silvicultural management.